Problem 1 Image Registration in the Presence
of Discontinuities(presentation)
Presenters: Yonho Kim, Dustin Steinhauer, Karteek Popuri, Rolf Clackdoyle,
Alvin Ihsani, Khaled Issa, Evgeniy Lebed
Image registration is one of the challenging problems in image processing
and specifically medical imaging. Given two images taken, for example,
at different times, and from different devices or perspectives,
the goal is to determine a reasonable geometric transformation that
aligns two images in a common frame of reference (full
abstract).
Reference: Brain-Tumor Interaction Biophysical
Models for Medical Image Registration
Deformable
Registration of Brain Tumor Images
Problem 2. Rapid modeling of internal structures
of deformable organs (i.e. liver) (presentation)
Presenters: Bahram Marami, Iain Moyles
Edward Xishi Huang and James Drake
CIGITI, The Hospital for Sick Children
Accurate estimation of deformation of soft organ internal structures
between two images acquired at different conditions. Boundary conditions:
vessel curves and point landmarks
Background:
Cancer is one of the leading causes of death in the developed countries
and the only major disease for which death rates are increasing.
For example, the number of children who suffer from tumors has been
increasing by about 0.6% per year. Although uncommon, cancer is
the second leading cause of death in children.
Recent development of high intensity focused ultrasound (HIFU) thermal
procedures has a great potential to provide a noninvasive alternative
solution to tumor treatment, which is expected to replace many current
invasive procedures for kids such as liver and kidney cancers. Although
HIFU procedures provide advantages such as less trauma and side
effects, their effectiveness are largely limited by the loss of
direct vision. During the procedure, doctors have to rely on real-time
image guidance for targeting treatment sites, sparing nearby normal
tissues, and monitoring thermal dose delivery. Currently, real-time
image guidance is limited by intra-operative low quality image and/or
slow image acquisition process. High quality real-time image
guidance is very challenging especially for deformable moving targets
such as the liver, which involve large organ shift and deformation,
and require dynamically steering HIFU beam to moving targets to
improve treatment efficiency.
In order to improve image guidance, we need to map pre-treatment
patient-specific deformable models and treatment plans to the patient
to complement limited real-time imaging. Therefore, accurate deformation
models of soft organs are an essential component for treatment planning,
treatment delivery, and assessment of disease progression/regression.
Problem: Rapid modeling of internal structures of deformable
organs (i.e. liver)
Our goal is to rapidly obtain 2D/3D models of deformable organs
by estimating deformation using two images acquired at different
conditions subject to multiple types of boundary conditions such
as vessel curves and point landmarks. This problem may be formulated
as deformation of elastic solid body, which can be solved by finite
element methods (FEM).
Finite element method is one of deformable image registration techniques,
aiming to find optimal geometric transformation of homologous points
in two images. A typical FEM problem in deformable image registration
is as follows: given tissue properties and boundary conditions extracted
from two images, FEM optimizes an objective function such as energy
in order to find the displacement/deformation field subject to boundary
conditions.
However, due to measurement errors in tissue properties and boundary
conditions, the accuracy and speed of conventional FEM methods is
often limited when used for deformable image registration of two
images, particularly for internal structures deep inside the organ.
To address these problems, we need boundary conditions to include
deep internal structure features such as centerlines of blood vessels
and bifurcations of vessels. That is, we need different types of
boundary conditions (i.e. curves of vessel centerlines and points
of vessel bifurcations) to constrain the solution. This will ensure
accurate alignment of internal structures deep inside organs.
After the curves of vessel centerlines are extracted from medical
images, they can be represented by smooth differential B-Spline
curves if it facilitates to solve this modeling problem. Extracted
vessel trees can be further processed into curve segments with known
end points.
One potential solution is that this deformation modeling can be
formulated as a minimum energy problem subject to vessel curves
and point landmarks.
Scenarios:
Problem 1a: Modeling of internal structures with 2D images
using vessel segments with both known end points.
Problem 1b: Same as 1 except only one known end point,
free for another end point, i.e. it is possibly not equal for
vessel lengths from two images due to imaging artifacts and extraction
errors.
Problem 2: Modeling of internal structures with 3D images
using vessel segments. It is easier to obtain more accurate curves
and point landmarks from 3D images for validation in human subjects
(2D problem is a simplified version).
Problem 3: Rapid 3D/4D modeling of deformable moving organs
(i.e. liver)
Simultaneous estimation of deformation and tissue properties using
multiple images acquired at different conditions
Boundary conditions: organ surface, 3D curves and 3D point landmarks
Acquisition of boundary conditions:
- Vessel centerlines: extracted from MR/CT images
- Branch points of vessels: extracted from MR/CT images
- (Organ surfaces: Segmented from MR/CT images or reconstructed
from stereo video images for Problem 3.)
Impact:
Accurate modeling of deformation of internal structures will improve
the quality of image guided treatment procedures deep inside soft
organs as these vessel structures provide a good reference to localize
deep treatment targets.
After accurate modeling of deformation of internal structures, modeling
of the whole deformable organ can be relatively easier to achieve
using other techniques.
Accurate modeling of deformation of internal structures between
two images acquired at different conditions will also be a fundamental
step towards accurate 3D/4D modeling of deformable moving organs
and intra-procedural image fusion.
Accurate models of deformable moving organs can facilitate treatment
planning and map high quality patient-specific models/images to
the patient. These models complement intra-operative images in order
to improve outcomes of image guided treatment procedures.
Problem 3 Detecting current density vector coherent
movement
Cerebral Diagnostics Canada Inc. (presentation.ppt)
Presenters: Sujanthan Sriskandarajah, Nataliya Portman, Alexandre
Foucault, Dominique Brunet, Yousef Akhavan, Vavara Nika (abstract)
I will describe the problem first in non-mathematical terms to
clarify what we are trying to achieve and why. Then I will attempt
to provide a mathematical frame work.
We want to be able to isolate current density signal patterns extracted
from (electroencephalography) EEG and to measure small transient
(often less than one quarter of one second) mental events in the
brain such as little cognitions like what you mind does when you
imagine the shape of a letter in the alphabet in your mind. If successful
there are numerous ramifications for this for neuroscience. For
example, it could help people communicate.
I, and many others, believe that these signals are hidden amongst
the many EEG signals taken while a person performs the cognition.
EEG itself plots voltage against time and, on its own, it is far
too crude to find the type signals we are looking for. I believe
this is because they are weak signals of low voltage that are coming
from tiny areas of the brain and they are overshadowed (buried)
among various larger signals.) For example it is known that there
are special areas of the brain in which language aspects of vision
and lettersare likely imagined.
We do source localization using an existing algorithm called eLoreta.
This gives us a data set at each instant in time providing4 numbers
for each voxel for each instant in time. The numbers are the x,
y and z components and magnitude of the current density for each
voxel. The voxels are in known positions.
When we make brain movies of these vectors we often see them dancing
together. We use our imaging software to draw the vectors as lines
radiating from the centre of each voxel. We can clearly see clusters
of voxels. For example, voxels 2 and 3 are side by side at the bottom
of the brain. In a given EEG recording they may be seen pivoting
in unison about the midpoints of the voxels. At a given time instant
within a given frequency band (e.g. 2-4 Hz brain activity) they
may be seen pivoting about their center points in near perfect synchronicity
the same movement
pattern. (This is a little hard to explain but easy to see in the
movies.) Hence as the vector radiating out of one voxel shifts in
the direction it is pointing, we might see the vector radiating
out of the adjacent voxel shifting in almost exactly the same way.
What we need in the long run is a real time tool to colour code
clusters and to make a list of the voxels that are members of a
cluster at an instant in time. For example if there are 10 voxels
in one cluster that are pivoting in one pattern then we need a colour
assigned to all the vectors in group.
In the short run, before tackling the issue of real time, we need
a post-processing method. Eventually we want this to be implemented
in C++ because we want to make it part of our brain movie software
bundle called DECI which is written in C++ and open GL.
Deci (Dynamic Electrical Cortical Imaging) is our software bundle.
It can be made available through a free software research license.
Data sets and sample movies are available to team members. This
can be used to validate any new tool created by the team. If your
tool works it will pick out the clusters of vectors we can see with
our own eyes dancing together in the movies played with DECI.
Ideally, the math team needs to work in close association with
computer programmers so that the project can culminate in a useful
software tool that functions well and is well explained. In the
future, once it works we plan to test it out by having people perform
small cognitions, and they
seeing if the software can help us find things such as a cluster
of vectors dancing together that are responsible for the cognition.
This ambitious goal, if achieved, would be a great advance for neuroscience.
Mark Doidge MD, Aug. 1, 2012 and amended Aug. 6 2012
(markdoidge@cerebraldiagnostics.com)
Problem 4 - Statistical models with tolerance
for abnormalities
Shuo Li, GE Healthcare
(download) (presentation)
Presentations: Craig Sinnamon, Anna Belkine, Berardo Galvao-Sousa
References
[1] M.M. Chakravarty,
G. Bertrand, C.P. Hodge, A.F. Sadikot, and D.L. Collins, The creation
of a brain atlas for image guided neurosurgery using serial histological
data, Neuroimage 30 (2006), no. 2, 359-376.
[2] X. Zhou, T. Kitagawa,
T. Hara, H. Fujita, X. Zhang, R. Yokoyama, H. Kondo, M. Kanematsu,
and H. Hoshi, Constructing a probabilistic model for automated liver
region segmentation using non-contrast x-ray torso ct images, Medical
Image Computing and Computer-Assisted Intervention{MICCAI 2006 (2006),
856-863.
Problem 5 - Modelling human perception in clinical diagnosis
Shuo Li, GE Healthcare Shuo.li@ge.com (download)
References
[1] G. Kong, D.L. Xu, and J.B.
Yang, Clinical decision support systems: a review on knowledge representation
and inference under uncertainties, International Journal of Computational
Intelligence Systems 1 (2008), no. 2, 159-167
[2] V.M.C.A. Van Belle, B.
Van Calster, D. Timmerman, T. Bourne, C. Bottomley, L. Valentin,
P. Neven, S. Van Hu el, J.A.K. Suykens, and S. Boyd, A mathematical
model for interpretable clinical
Back to Top
